vfs_bio.c revision 18271
1/* 2 * Copyright (c) 1994 John S. Dyson 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice immediately at the beginning of the file, without modification, 10 * this list of conditions, and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 3. Absolutely no warranty of function or purpose is made by the author 15 * John S. Dyson. 16 * 4. This work was done expressly for inclusion into FreeBSD. Other use 17 * is allowed if this notation is included. 18 * 5. Modifications may be freely made to this file if the above conditions 19 * are met. 20 * 21 * $Id: vfs_bio.c,v 1.98 1996/09/08 20:44:20 dyson Exp $ 22 */ 23 24/* 25 * this file contains a new buffer I/O scheme implementing a coherent 26 * VM object and buffer cache scheme. Pains have been taken to make 27 * sure that the performance degradation associated with schemes such 28 * as this is not realized. 29 * 30 * Author: John S. Dyson 31 * Significant help during the development and debugging phases 32 * had been provided by David Greenman, also of the FreeBSD core team. 33 */ 34 35#include "opt_bounce.h" 36 37#define VMIO 38#include <sys/param.h> 39#include <sys/systm.h> 40#include <sys/sysproto.h> 41#include <sys/kernel.h> 42#include <sys/sysctl.h> 43#include <sys/proc.h> 44#include <sys/vnode.h> 45#include <sys/vmmeter.h> 46#include <vm/vm.h> 47#include <vm/vm_param.h> 48#include <vm/vm_prot.h> 49#include <vm/vm_kern.h> 50#include <vm/vm_pageout.h> 51#include <vm/vm_page.h> 52#include <vm/vm_object.h> 53#include <vm/vm_extern.h> 54#include <sys/buf.h> 55#include <sys/mount.h> 56#include <sys/malloc.h> 57#include <sys/resourcevar.h> 58#include <sys/proc.h> 59 60#include <miscfs/specfs/specdev.h> 61 62static void vfs_update __P((void)); 63static struct proc *updateproc; 64static struct kproc_desc up_kp = { 65 "update", 66 vfs_update, 67 &updateproc 68}; 69SYSINIT_KT(update, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp) 70 71struct buf *buf; /* buffer header pool */ 72struct swqueue bswlist; 73 74int count_lock_queue __P((void)); 75static void vm_hold_free_pages(struct buf * bp, vm_offset_t from, 76 vm_offset_t to); 77static void vm_hold_load_pages(struct buf * bp, vm_offset_t from, 78 vm_offset_t to); 79static void vfs_clean_pages(struct buf * bp); 80static void vfs_setdirty(struct buf *bp); 81static void vfs_vmio_release(struct buf *bp); 82 83int needsbuffer; 84 85/* 86 * Internal update daemon, process 3 87 * The variable vfs_update_wakeup allows for internal syncs. 88 */ 89int vfs_update_wakeup; 90 91 92/* 93 * buffers base kva 94 */ 95caddr_t buffers_kva; 96 97/* 98 * bogus page -- for I/O to/from partially complete buffers 99 * this is a temporary solution to the problem, but it is not 100 * really that bad. it would be better to split the buffer 101 * for input in the case of buffers partially already in memory, 102 * but the code is intricate enough already. 103 */ 104vm_page_t bogus_page; 105static vm_offset_t bogus_offset; 106 107static int bufspace, maxbufspace, vmiospace, maxvmiobufspace, 108 bufmallocspace, maxbufmallocspace; 109 110static struct bufhashhdr bufhashtbl[BUFHSZ], invalhash; 111static struct bqueues bufqueues[BUFFER_QUEUES]; 112 113extern int vm_swap_size; 114 115#define BUF_MAXUSE 8 116/* 117#define NO_B_MALLOC 118*/ 119 120/* 121 * Initialize buffer headers and related structures. 122 */ 123void 124bufinit() 125{ 126 struct buf *bp; 127 int i; 128 129 TAILQ_INIT(&bswlist); 130 LIST_INIT(&invalhash); 131 132 /* first, make a null hash table */ 133 for (i = 0; i < BUFHSZ; i++) 134 LIST_INIT(&bufhashtbl[i]); 135 136 /* next, make a null set of free lists */ 137 for (i = 0; i < BUFFER_QUEUES; i++) 138 TAILQ_INIT(&bufqueues[i]); 139 140 buffers_kva = (caddr_t) kmem_alloc_pageable(buffer_map, MAXBSIZE * nbuf); 141 /* finally, initialize each buffer header and stick on empty q */ 142 for (i = 0; i < nbuf; i++) { 143 bp = &buf[i]; 144 bzero(bp, sizeof *bp); 145 bp->b_flags = B_INVAL; /* we're just an empty header */ 146 bp->b_dev = NODEV; 147 bp->b_rcred = NOCRED; 148 bp->b_wcred = NOCRED; 149 bp->b_qindex = QUEUE_EMPTY; 150 bp->b_vnbufs.le_next = NOLIST; 151 bp->b_data = buffers_kva + i * MAXBSIZE; 152 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 153 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 154 } 155/* 156 * maxbufspace is currently calculated to support all filesystem blocks 157 * to be 8K. If you happen to use a 16K filesystem, the size of the buffer 158 * cache is still the same as it would be for 8K filesystems. This 159 * keeps the size of the buffer cache "in check" for big block filesystems. 160 */ 161 maxbufspace = 2 * (nbuf + 8) * PAGE_SIZE; 162/* 163 * reserve 1/3 of the buffers for metadata (VDIR) which might not be VMIO'ed 164 */ 165 maxvmiobufspace = 2 * maxbufspace / 3; 166/* 167 * Limit the amount of malloc memory since it is wired permanently into 168 * the kernel space. Even though this is accounted for in the buffer 169 * allocation, we don't want the malloced region to grow uncontrolled. 170 * The malloc scheme improves memory utilization significantly on average 171 * (small) directories. 172 */ 173 maxbufmallocspace = maxbufspace / 20; 174 175 bogus_offset = kmem_alloc_pageable(kernel_map, PAGE_SIZE); 176 bogus_page = vm_page_alloc(kernel_object, 177 ((bogus_offset - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 178 VM_ALLOC_NORMAL); 179 180} 181 182/* 183 * remove the buffer from the appropriate free list 184 */ 185void 186bremfree(struct buf * bp) 187{ 188 int s = splbio(); 189 190 if (bp->b_qindex != QUEUE_NONE) { 191 TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist); 192 bp->b_qindex = QUEUE_NONE; 193 } else { 194 panic("bremfree: removing a buffer when not on a queue"); 195 } 196 splx(s); 197} 198 199/* 200 * Get a buffer with the specified data. Look in the cache first. 201 */ 202int 203bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred, 204 struct buf ** bpp) 205{ 206 struct buf *bp; 207 208 bp = getblk(vp, blkno, size, 0, 0); 209 *bpp = bp; 210 211 /* if not found in cache, do some I/O */ 212 if ((bp->b_flags & B_CACHE) == 0) { 213 if (curproc != NULL) 214 curproc->p_stats->p_ru.ru_inblock++; 215 bp->b_flags |= B_READ; 216 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 217 if (bp->b_rcred == NOCRED) { 218 if (cred != NOCRED) 219 crhold(cred); 220 bp->b_rcred = cred; 221 } 222 vfs_busy_pages(bp, 0); 223 VOP_STRATEGY(bp); 224 return (biowait(bp)); 225 } 226 return (0); 227} 228 229/* 230 * Operates like bread, but also starts asynchronous I/O on 231 * read-ahead blocks. 232 */ 233int 234breadn(struct vnode * vp, daddr_t blkno, int size, 235 daddr_t * rablkno, int *rabsize, 236 int cnt, struct ucred * cred, struct buf ** bpp) 237{ 238 struct buf *bp, *rabp; 239 int i; 240 int rv = 0, readwait = 0; 241 242 *bpp = bp = getblk(vp, blkno, size, 0, 0); 243 244 /* if not found in cache, do some I/O */ 245 if ((bp->b_flags & B_CACHE) == 0) { 246 if (curproc != NULL) 247 curproc->p_stats->p_ru.ru_inblock++; 248 bp->b_flags |= B_READ; 249 bp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 250 if (bp->b_rcred == NOCRED) { 251 if (cred != NOCRED) 252 crhold(cred); 253 bp->b_rcred = cred; 254 } 255 vfs_busy_pages(bp, 0); 256 VOP_STRATEGY(bp); 257 ++readwait; 258 } 259 for (i = 0; i < cnt; i++, rablkno++, rabsize++) { 260 if (inmem(vp, *rablkno)) 261 continue; 262 rabp = getblk(vp, *rablkno, *rabsize, 0, 0); 263 264 if ((rabp->b_flags & B_CACHE) == 0) { 265 if (curproc != NULL) 266 curproc->p_stats->p_ru.ru_inblock++; 267 rabp->b_flags |= B_READ | B_ASYNC; 268 rabp->b_flags &= ~(B_DONE | B_ERROR | B_INVAL); 269 if (rabp->b_rcred == NOCRED) { 270 if (cred != NOCRED) 271 crhold(cred); 272 rabp->b_rcred = cred; 273 } 274 vfs_busy_pages(rabp, 0); 275 VOP_STRATEGY(rabp); 276 } else { 277 brelse(rabp); 278 } 279 } 280 281 if (readwait) { 282 rv = biowait(bp); 283 } 284 return (rv); 285} 286 287/* 288 * Write, release buffer on completion. (Done by iodone 289 * if async.) 290 */ 291int 292bwrite(struct buf * bp) 293{ 294 int oldflags = bp->b_flags; 295 296 if (bp->b_flags & B_INVAL) { 297 brelse(bp); 298 return (0); 299 } 300 if (!(bp->b_flags & B_BUSY)) 301 panic("bwrite: buffer is not busy???"); 302 303 bp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI); 304 bp->b_flags |= B_WRITEINPROG; 305 306 if ((oldflags & (B_ASYNC|B_DELWRI)) == (B_ASYNC|B_DELWRI)) { 307 reassignbuf(bp, bp->b_vp); 308 } 309 310 bp->b_vp->v_numoutput++; 311 vfs_busy_pages(bp, 1); 312 if (curproc != NULL) 313 curproc->p_stats->p_ru.ru_oublock++; 314 VOP_STRATEGY(bp); 315 316 /* if ((bp->b_flags & B_ASYNC) == 0) { */ 317 if ((oldflags & B_ASYNC) == 0) { 318 int rtval = biowait(bp); 319 320 if (oldflags & B_DELWRI) { 321 reassignbuf(bp, bp->b_vp); 322 } 323 brelse(bp); 324 return (rtval); 325 } 326 return (0); 327} 328 329int 330vn_bwrite(ap) 331 struct vop_bwrite_args *ap; 332{ 333 return (bwrite(ap->a_bp)); 334} 335 336/* 337 * Delayed write. (Buffer is marked dirty). 338 */ 339void 340bdwrite(struct buf * bp) 341{ 342 343 if ((bp->b_flags & B_BUSY) == 0) { 344 panic("bdwrite: buffer is not busy"); 345 } 346 if (bp->b_flags & B_INVAL) { 347 brelse(bp); 348 return; 349 } 350 if (bp->b_flags & B_TAPE) { 351 bawrite(bp); 352 return; 353 } 354 bp->b_flags &= ~(B_READ|B_RELBUF); 355 if ((bp->b_flags & B_DELWRI) == 0) { 356 bp->b_flags |= B_DONE | B_DELWRI; 357 reassignbuf(bp, bp->b_vp); 358 } 359 360 /* 361 * This bmap keeps the system from needing to do the bmap later, 362 * perhaps when the system is attempting to do a sync. Since it 363 * is likely that the indirect block -- or whatever other datastructure 364 * that the filesystem needs is still in memory now, it is a good 365 * thing to do this. Note also, that if the pageout daemon is 366 * requesting a sync -- there might not be enough memory to do 367 * the bmap then... So, this is important to do. 368 */ 369 if( bp->b_lblkno == bp->b_blkno) { 370 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); 371 } 372 373 /* 374 * Set the *dirty* buffer range based upon the VM system dirty pages. 375 */ 376 vfs_setdirty(bp); 377 378 /* 379 * We need to do this here to satisfy the vnode_pager and the 380 * pageout daemon, so that it thinks that the pages have been 381 * "cleaned". Note that since the pages are in a delayed write 382 * buffer -- the VFS layer "will" see that the pages get written 383 * out on the next sync, or perhaps the cluster will be completed. 384 */ 385 vfs_clean_pages(bp); 386 bqrelse(bp); 387 return; 388} 389 390/* 391 * Asynchronous write. 392 * Start output on a buffer, but do not wait for it to complete. 393 * The buffer is released when the output completes. 394 */ 395void 396bawrite(struct buf * bp) 397{ 398 bp->b_flags |= B_ASYNC; 399 (void) VOP_BWRITE(bp); 400} 401 402/* 403 * Ordered write. 404 * Start output on a buffer, but only wait for it to complete if the 405 * output device cannot guarantee ordering in some other way. Devices 406 * that can perform asynchronous ordered writes will set the B_ASYNC 407 * flag in their strategy routine. 408 * The buffer is released when the output completes. 409 */ 410int 411bowrite(struct buf * bp) 412{ 413 bp->b_flags |= B_ORDERED; 414 return (VOP_BWRITE(bp)); 415} 416 417/* 418 * Release a buffer. 419 */ 420void 421brelse(struct buf * bp) 422{ 423 int s; 424 425 if (bp->b_flags & B_CLUSTER) { 426 relpbuf(bp); 427 return; 428 } 429 /* anyone need a "free" block? */ 430 s = splbio(); 431 432 /* anyone need this block? */ 433 if (bp->b_flags & B_WANTED) { 434 bp->b_flags &= ~(B_WANTED | B_AGE); 435 wakeup(bp); 436 } 437 438 if (bp->b_flags & B_LOCKED) 439 bp->b_flags &= ~B_ERROR; 440 441 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) || 442 (bp->b_bufsize <= 0)) { 443 bp->b_flags |= B_INVAL; 444 bp->b_flags &= ~(B_DELWRI | B_CACHE); 445 if (((bp->b_flags & B_VMIO) == 0) && bp->b_vp) { 446 if (bp->b_bufsize) 447 allocbuf(bp, 0); 448 brelvp(bp); 449 } 450 } 451 452 /* 453 * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer 454 * constituted, so the B_INVAL flag is used to *invalidate* the buffer, 455 * but the VM object is kept around. The B_NOCACHE flag is used to 456 * invalidate the pages in the VM object. 457 */ 458 if (bp->b_flags & B_VMIO) { 459 vm_ooffset_t foff; 460 vm_object_t obj; 461 int i, resid; 462 vm_page_t m; 463 struct vnode *vp; 464 int iototal = bp->b_bufsize; 465 466 vp = bp->b_vp; 467 if (!vp) 468 panic("brelse: missing vp"); 469 470 if (bp->b_npages) { 471 vm_pindex_t poff; 472 obj = (vm_object_t) vp->v_object; 473 if (vp->v_type == VBLK) 474 foff = ((vm_ooffset_t) bp->b_lblkno) << DEV_BSHIFT; 475 else 476 foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 477 poff = OFF_TO_IDX(foff); 478 for (i = 0; i < bp->b_npages; i++) { 479 m = bp->b_pages[i]; 480 if (m == bogus_page) { 481 m = vm_page_lookup(obj, poff + i); 482 if (!m) { 483 panic("brelse: page missing\n"); 484 } 485 bp->b_pages[i] = m; 486 pmap_qenter(trunc_page(bp->b_data), 487 bp->b_pages, bp->b_npages); 488 } 489 resid = IDX_TO_OFF(m->pindex+1) - foff; 490 if (resid > iototal) 491 resid = iototal; 492 if (resid > 0) { 493 /* 494 * Don't invalidate the page if the local machine has already 495 * modified it. This is the lesser of two evils, and should 496 * be fixed. 497 */ 498 if (bp->b_flags & (B_NOCACHE | B_ERROR)) { 499 vm_page_test_dirty(m); 500 if (m->dirty == 0) { 501 vm_page_set_invalid(m, (vm_offset_t) foff, resid); 502 if (m->valid == 0) 503 vm_page_protect(m, VM_PROT_NONE); 504 } 505 } 506 if (resid >= PAGE_SIZE) { 507 if ((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) { 508 bp->b_flags |= B_INVAL; 509 } 510 } else { 511 if (!vm_page_is_valid(m, 512 (((vm_offset_t) bp->b_data) & PAGE_MASK), resid)) { 513 bp->b_flags |= B_INVAL; 514 } 515 } 516 } 517 foff += resid; 518 iototal -= resid; 519 } 520 } 521 if (bp->b_flags & (B_INVAL | B_RELBUF)) 522 vfs_vmio_release(bp); 523 } 524 if (bp->b_qindex != QUEUE_NONE) 525 panic("brelse: free buffer onto another queue???"); 526 527 /* enqueue */ 528 /* buffers with no memory */ 529 if (bp->b_bufsize == 0) { 530 bp->b_qindex = QUEUE_EMPTY; 531 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_EMPTY], bp, b_freelist); 532 LIST_REMOVE(bp, b_hash); 533 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 534 bp->b_dev = NODEV; 535 if (needsbuffer) { 536 wakeup(&needsbuffer); 537 needsbuffer=0; 538 } 539 /* buffers with junk contents */ 540 } else if (bp->b_flags & (B_ERROR | B_INVAL | B_NOCACHE | B_RELBUF)) { 541 bp->b_qindex = QUEUE_AGE; 542 TAILQ_INSERT_HEAD(&bufqueues[QUEUE_AGE], bp, b_freelist); 543 LIST_REMOVE(bp, b_hash); 544 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 545 bp->b_dev = NODEV; 546 if (needsbuffer) { 547 wakeup(&needsbuffer); 548 needsbuffer=0; 549 } 550 /* buffers that are locked */ 551 } else if (bp->b_flags & B_LOCKED) { 552 bp->b_qindex = QUEUE_LOCKED; 553 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 554 /* buffers with stale but valid contents */ 555 } else if (bp->b_flags & B_AGE) { 556 bp->b_qindex = QUEUE_AGE; 557 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_AGE], bp, b_freelist); 558 if (needsbuffer) { 559 wakeup(&needsbuffer); 560 needsbuffer=0; 561 } 562 /* buffers with valid and quite potentially reuseable contents */ 563 } else { 564 bp->b_qindex = QUEUE_LRU; 565 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 566 if (needsbuffer) { 567 wakeup(&needsbuffer); 568 needsbuffer=0; 569 } 570 } 571 572 /* unlock */ 573 bp->b_flags &= ~(B_WANTED | B_BUSY | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 574 splx(s); 575} 576 577/* 578 * Release a buffer. 579 */ 580void 581bqrelse(struct buf * bp) 582{ 583 int s; 584 585 s = splbio(); 586 587 588 /* anyone need this block? */ 589 if (bp->b_flags & B_WANTED) { 590 bp->b_flags &= ~(B_WANTED | B_AGE); 591 wakeup(bp); 592 } 593 594 if (bp->b_qindex != QUEUE_NONE) 595 panic("bqrelse: free buffer onto another queue???"); 596 597 if (bp->b_flags & B_LOCKED) { 598 bp->b_flags &= ~B_ERROR; 599 bp->b_qindex = QUEUE_LOCKED; 600 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LOCKED], bp, b_freelist); 601 /* buffers with stale but valid contents */ 602 } else { 603 bp->b_qindex = QUEUE_LRU; 604 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 605 if (needsbuffer) { 606 wakeup(&needsbuffer); 607 needsbuffer=0; 608 } 609 } 610 611 /* unlock */ 612 bp->b_flags &= ~(B_WANTED | B_BUSY | B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); 613 splx(s); 614} 615 616static void 617vfs_vmio_release(bp) 618 struct buf *bp; 619{ 620 int i; 621 vm_page_t m; 622 623 for (i = 0; i < bp->b_npages; i++) { 624 m = bp->b_pages[i]; 625 bp->b_pages[i] = NULL; 626 if (m->flags & PG_WANTED) { 627 m->flags &= ~PG_WANTED; 628 wakeup(m); 629 } 630 vm_page_unwire(m); 631 if (m->wire_count == 0 && (m->flags & PG_BUSY) == 0) { 632 if (m->valid) { 633 if(m->dirty == 0) 634 vm_page_test_dirty(m); 635 /* 636 * this keeps pressure off of the process memory 637 */ 638 if ((vm_swap_size == 0) || 639 (cnt.v_free_count < cnt.v_free_min)) { 640 if ((m->dirty == 0) && 641 (m->hold_count == 0) && 642 (m->flags & PG_BUSY) == 0 && 643 (m->busy == 0)) 644 vm_page_cache(m); 645 else 646 vm_page_deactivate(m); 647 } 648 } else if ((m->hold_count == 0) && 649 ((m->flags & PG_BUSY) == 0) && 650 (m->busy == 0)) { 651 vm_page_protect(m, VM_PROT_NONE); 652 vm_page_free(m); 653 } 654 } 655 } 656 bufspace -= bp->b_bufsize; 657 vmiospace -= bp->b_bufsize; 658 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); 659 bp->b_npages = 0; 660 bp->b_bufsize = 0; 661 bp->b_flags &= ~B_VMIO; 662 if (bp->b_vp) 663 brelvp(bp); 664} 665 666/* 667 * Check to see if a block is currently memory resident. 668 */ 669__inline struct buf * 670gbincore(struct vnode * vp, daddr_t blkno) 671{ 672 struct buf *bp; 673 struct bufhashhdr *bh; 674 675 bh = BUFHASH(vp, blkno); 676 bp = bh->lh_first; 677 678 /* Search hash chain */ 679 while (bp != NULL) { 680 /* hit */ 681 if (bp->b_vp == vp && bp->b_lblkno == blkno && 682 (bp->b_flags & B_INVAL) == 0) { 683 break; 684 } 685 bp = bp->b_hash.le_next; 686 } 687 return (bp); 688} 689 690/* 691 * this routine implements clustered async writes for 692 * clearing out B_DELWRI buffers... This is much better 693 * than the old way of writing only one buffer at a time. 694 */ 695int 696vfs_bio_awrite(struct buf * bp) 697{ 698 int i; 699 daddr_t lblkno = bp->b_lblkno; 700 struct vnode *vp = bp->b_vp; 701 int s; 702 int ncl; 703 struct buf *bpa; 704 int nwritten; 705 706 s = splbio(); 707 /* 708 * right now we support clustered writing only to regular files 709 */ 710 if ((vp->v_type == VREG) && 711 (vp->v_mount != 0) && /* Only on nodes that have the size info */ 712 (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { 713 int size; 714 int maxcl; 715 716 size = vp->v_mount->mnt_stat.f_iosize; 717 maxcl = MAXPHYS / size; 718 719 for (i = 1; i < maxcl; i++) { 720 if ((bpa = gbincore(vp, lblkno + i)) && 721 ((bpa->b_flags & (B_BUSY | B_DELWRI | B_CLUSTEROK | B_INVAL)) == 722 (B_DELWRI | B_CLUSTEROK)) && 723 (bpa->b_bufsize == size)) { 724 if ((bpa->b_blkno == bpa->b_lblkno) || 725 (bpa->b_blkno != bp->b_blkno + ((i * size) >> DEV_BSHIFT))) 726 break; 727 } else { 728 break; 729 } 730 } 731 ncl = i; 732 /* 733 * this is a possible cluster write 734 */ 735 if (ncl != 1) { 736 nwritten = cluster_wbuild(vp, size, lblkno, ncl); 737 splx(s); 738 return nwritten; 739 } 740 } 741 bremfree(bp); 742 splx(s); 743 /* 744 * default (old) behavior, writing out only one block 745 */ 746 bp->b_flags |= B_BUSY | B_ASYNC; 747 nwritten = bp->b_bufsize; 748 (void) VOP_BWRITE(bp); 749 return nwritten; 750} 751 752 753/* 754 * Find a buffer header which is available for use. 755 */ 756static struct buf * 757getnewbuf(int slpflag, int slptimeo, int doingvmio) 758{ 759 struct buf *bp; 760 int nbyteswritten = 0; 761 762start: 763 if (bufspace >= maxbufspace) 764 goto trytofreespace; 765 766 /* can we constitute a new buffer? */ 767 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_EMPTY]))) { 768 if (bp->b_qindex != QUEUE_EMPTY) 769 panic("getnewbuf: inconsistent EMPTY queue, qindex=%d", 770 bp->b_qindex); 771 bp->b_flags |= B_BUSY; 772 bremfree(bp); 773 goto fillbuf; 774 } 775trytofreespace: 776 /* 777 * We keep the file I/O from hogging metadata I/O 778 * This is desirable because file data is cached in the 779 * VM/Buffer cache even if a buffer is freed. 780 */ 781 if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_AGE]))) { 782 if (bp->b_qindex != QUEUE_AGE) 783 panic("getnewbuf: inconsistent AGE queue, qindex=%d", 784 bp->b_qindex); 785 } else if ((bp = TAILQ_FIRST(&bufqueues[QUEUE_LRU]))) { 786 if (bp->b_qindex != QUEUE_LRU) 787 panic("getnewbuf: inconsistent LRU queue, qindex=%d", 788 bp->b_qindex); 789 } 790 if (!bp) { 791 /* wait for a free buffer of any kind */ 792 needsbuffer = 1; 793 tsleep(&needsbuffer, 794 (PRIBIO + 1) | slpflag, "newbuf", slptimeo); 795 return (0); 796 } 797 798 /* 799 * We are fairly aggressive about freeing VMIO buffers, but since 800 * the buffering is intact without buffer headers, there is not 801 * much loss. We gain by maintaining non-VMIOed metadata in buffers. 802 */ 803 if ((bp->b_qindex == QUEUE_LRU) && (bp->b_usecount > 0)) { 804 if ((bp->b_flags & B_VMIO) == 0 || 805 (vmiospace < maxvmiobufspace)) { 806 --bp->b_usecount; 807 TAILQ_REMOVE(&bufqueues[QUEUE_LRU], bp, b_freelist); 808 if (TAILQ_FIRST(&bufqueues[QUEUE_LRU]) != NULL) { 809 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 810 goto start; 811 } 812 TAILQ_INSERT_TAIL(&bufqueues[QUEUE_LRU], bp, b_freelist); 813 } 814 } 815 816 /* if we are a delayed write, convert to an async write */ 817 if ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) { 818 nbyteswritten += vfs_bio_awrite(bp); 819 if (!slpflag && !slptimeo) { 820 return (0); 821 } 822 goto start; 823 } 824 825 if (bp->b_flags & B_WANTED) { 826 bp->b_flags &= ~B_WANTED; 827 wakeup(bp); 828 } 829 bremfree(bp); 830 bp->b_flags |= B_BUSY; 831 832 if (bp->b_flags & B_VMIO) 833 vfs_vmio_release(bp); 834 835 if (bp->b_vp) 836 brelvp(bp); 837 838fillbuf: 839 /* we are not free, nor do we contain interesting data */ 840 if (bp->b_rcred != NOCRED) { 841 crfree(bp->b_rcred); 842 bp->b_rcred = NOCRED; 843 } 844 if (bp->b_wcred != NOCRED) { 845 crfree(bp->b_wcred); 846 bp->b_wcred = NOCRED; 847 } 848 849 LIST_REMOVE(bp, b_hash); 850 LIST_INSERT_HEAD(&invalhash, bp, b_hash); 851 if (bp->b_bufsize) { 852 allocbuf(bp, 0); 853 } 854 bp->b_flags = B_BUSY; 855 bp->b_dev = NODEV; 856 bp->b_vp = NULL; 857 bp->b_blkno = bp->b_lblkno = 0; 858 bp->b_iodone = 0; 859 bp->b_error = 0; 860 bp->b_resid = 0; 861 bp->b_bcount = 0; 862 bp->b_npages = 0; 863 bp->b_data = buffers_kva + (bp - buf) * MAXBSIZE; 864 bp->b_dirtyoff = bp->b_dirtyend = 0; 865 bp->b_validoff = bp->b_validend = 0; 866 bp->b_usecount = 4; 867 if (bufspace >= maxbufspace + nbyteswritten) { 868 bp->b_flags |= B_INVAL; 869 brelse(bp); 870 goto trytofreespace; 871 } 872 return (bp); 873} 874 875/* 876 * Check to see if a block is currently memory resident. 877 */ 878struct buf * 879incore(struct vnode * vp, daddr_t blkno) 880{ 881 struct buf *bp; 882 883 int s = splbio(); 884 bp = gbincore(vp, blkno); 885 splx(s); 886 return (bp); 887} 888 889/* 890 * Returns true if no I/O is needed to access the 891 * associated VM object. This is like incore except 892 * it also hunts around in the VM system for the data. 893 */ 894 895int 896inmem(struct vnode * vp, daddr_t blkno) 897{ 898 vm_object_t obj; 899 vm_offset_t toff, tinc; 900 vm_page_t m; 901 vm_ooffset_t off; 902 903 if (incore(vp, blkno)) 904 return 1; 905 if (vp->v_mount == NULL) 906 return 0; 907 if ((vp->v_object == NULL) || (vp->v_flag & VVMIO) == 0) 908 return 0; 909 910 obj = vp->v_object; 911 tinc = PAGE_SIZE; 912 if (tinc > vp->v_mount->mnt_stat.f_iosize) 913 tinc = vp->v_mount->mnt_stat.f_iosize; 914 off = blkno * vp->v_mount->mnt_stat.f_iosize; 915 916 for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { 917 918 m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); 919 if (!m) 920 return 0; 921 if (vm_page_is_valid(m, (vm_offset_t) (toff + off), tinc) == 0) 922 return 0; 923 } 924 return 1; 925} 926 927/* 928 * now we set the dirty range for the buffer -- 929 * for NFS -- if the file is mapped and pages have 930 * been written to, let it know. We want the 931 * entire range of the buffer to be marked dirty if 932 * any of the pages have been written to for consistancy 933 * with the b_validoff, b_validend set in the nfs write 934 * code, and used by the nfs read code. 935 */ 936static void 937vfs_setdirty(struct buf *bp) { 938 int i; 939 vm_object_t object; 940 vm_offset_t boffset, offset; 941 /* 942 * We qualify the scan for modified pages on whether the 943 * object has been flushed yet. The OBJ_WRITEABLE flag 944 * is not cleared simply by protecting pages off. 945 */ 946 if ((bp->b_flags & B_VMIO) && 947 ((object = bp->b_pages[0]->object)->flags & (OBJ_WRITEABLE|OBJ_CLEANING))) { 948 /* 949 * test the pages to see if they have been modified directly 950 * by users through the VM system. 951 */ 952 for (i = 0; i < bp->b_npages; i++) 953 vm_page_test_dirty(bp->b_pages[i]); 954 955 /* 956 * scan forwards for the first page modified 957 */ 958 for (i = 0; i < bp->b_npages; i++) { 959 if (bp->b_pages[i]->dirty) { 960 break; 961 } 962 } 963 boffset = (i << PAGE_SHIFT); 964 if (boffset < bp->b_dirtyoff) { 965 bp->b_dirtyoff = boffset; 966 } 967 968 /* 969 * scan backwards for the last page modified 970 */ 971 for (i = bp->b_npages - 1; i >= 0; --i) { 972 if (bp->b_pages[i]->dirty) { 973 break; 974 } 975 } 976 boffset = (i + 1); 977 offset = boffset + bp->b_pages[0]->pindex; 978 if (offset >= object->size) 979 boffset = object->size - bp->b_pages[0]->pindex; 980 if (bp->b_dirtyend < (boffset << PAGE_SHIFT)) 981 bp->b_dirtyend = (boffset << PAGE_SHIFT); 982 } 983} 984 985/* 986 * Get a block given a specified block and offset into a file/device. 987 */ 988struct buf * 989getblk(struct vnode * vp, daddr_t blkno, int size, int slpflag, int slptimeo) 990{ 991 struct buf *bp; 992 int s; 993 struct bufhashhdr *bh; 994 995 s = splbio(); 996loop: 997 if ((bp = gbincore(vp, blkno))) { 998 if (bp->b_flags & B_BUSY) { 999 bp->b_flags |= B_WANTED; 1000 if (bp->b_usecount < BUF_MAXUSE) 1001 ++bp->b_usecount; 1002 if (!tsleep(bp, 1003 (PRIBIO + 1) | slpflag, "getblk", slptimeo)) 1004 goto loop; 1005 1006 splx(s); 1007 return (struct buf *) NULL; 1008 } 1009 bp->b_flags |= B_BUSY | B_CACHE; 1010 bremfree(bp); 1011 1012 /* 1013 * check for size inconsistancies (note that they shouldn't happen 1014 * but do when filesystems don't handle the size changes correctly.) 1015 * We are conservative on metadata and don't just extend the buffer 1016 * but write and re-constitute it. 1017 */ 1018 1019 if (bp->b_bcount != size) { 1020 if (bp->b_flags & B_VMIO) { 1021 allocbuf(bp, size); 1022 } else { 1023 bp->b_flags |= B_NOCACHE; 1024 VOP_BWRITE(bp); 1025 goto loop; 1026 } 1027 } 1028 1029 if (bp->b_usecount < BUF_MAXUSE) 1030 ++bp->b_usecount; 1031 splx(s); 1032 return (bp); 1033 } else { 1034 vm_object_t obj; 1035 int doingvmio; 1036 1037 if ((obj = vp->v_object) && (vp->v_flag & VVMIO)) { 1038 doingvmio = 1; 1039 } else { 1040 doingvmio = 0; 1041 } 1042 if ((bp = getnewbuf(slpflag, slptimeo, doingvmio)) == 0) { 1043 if (slpflag || slptimeo) { 1044 splx(s); 1045 return NULL; 1046 } 1047 goto loop; 1048 } 1049 1050 /* 1051 * This code is used to make sure that a buffer is not 1052 * created while the getnewbuf routine is blocked. 1053 * Normally the vnode is locked so this isn't a problem. 1054 * VBLK type I/O requests, however, don't lock the vnode. 1055 */ 1056 if (!VOP_ISLOCKED(vp) && gbincore(vp, blkno)) { 1057 bp->b_flags |= B_INVAL; 1058 brelse(bp); 1059 goto loop; 1060 } 1061 1062 /* 1063 * Insert the buffer into the hash, so that it can 1064 * be found by incore. 1065 */ 1066 bp->b_blkno = bp->b_lblkno = blkno; 1067 bgetvp(vp, bp); 1068 LIST_REMOVE(bp, b_hash); 1069 bh = BUFHASH(vp, blkno); 1070 LIST_INSERT_HEAD(bh, bp, b_hash); 1071 1072 if (doingvmio) { 1073 bp->b_flags |= (B_VMIO | B_CACHE); 1074#if defined(VFS_BIO_DEBUG) 1075 if (vp->v_type != VREG && vp->v_type != VBLK) 1076 printf("getblk: vmioing file type %d???\n", vp->v_type); 1077#endif 1078 } else { 1079 bp->b_flags &= ~B_VMIO; 1080 } 1081 splx(s); 1082 1083 allocbuf(bp, size); 1084#ifdef PC98 1085 /* 1086 * 1024byte/sector support 1087 */ 1088#define B_XXX2 0x8000000 1089 if (vp->v_flag & 0x10000) bp->b_flags |= B_XXX2; 1090#endif 1091 return (bp); 1092 } 1093} 1094 1095/* 1096 * Get an empty, disassociated buffer of given size. 1097 */ 1098struct buf * 1099geteblk(int size) 1100{ 1101 struct buf *bp; 1102 int s; 1103 1104 s = splbio(); 1105 while ((bp = getnewbuf(0, 0, 0)) == 0); 1106 splx(s); 1107 allocbuf(bp, size); 1108 bp->b_flags |= B_INVAL; 1109 return (bp); 1110} 1111 1112 1113/* 1114 * This code constitutes the buffer memory from either anonymous system 1115 * memory (in the case of non-VMIO operations) or from an associated 1116 * VM object (in the case of VMIO operations). 1117 * 1118 * Note that this code is tricky, and has many complications to resolve 1119 * deadlock or inconsistant data situations. Tread lightly!!! 1120 * 1121 * Modify the length of a buffer's underlying buffer storage without 1122 * destroying information (unless, of course the buffer is shrinking). 1123 */ 1124int 1125allocbuf(struct buf * bp, int size) 1126{ 1127 1128 int s; 1129 int newbsize, mbsize; 1130 int i; 1131 1132 if (!(bp->b_flags & B_BUSY)) 1133 panic("allocbuf: buffer not busy"); 1134 1135 if ((bp->b_flags & B_VMIO) == 0) { 1136 caddr_t origbuf; 1137 int origbufsize; 1138 /* 1139 * Just get anonymous memory from the kernel 1140 */ 1141 mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1142#if !defined(NO_B_MALLOC) 1143 if (bp->b_flags & B_MALLOC) 1144 newbsize = mbsize; 1145 else 1146#endif 1147 newbsize = round_page(size); 1148 1149 if (newbsize < bp->b_bufsize) { 1150#if !defined(NO_B_MALLOC) 1151 /* 1152 * malloced buffers are not shrunk 1153 */ 1154 if (bp->b_flags & B_MALLOC) { 1155 if (newbsize) { 1156 bp->b_bcount = size; 1157 } else { 1158 free(bp->b_data, M_BIOBUF); 1159 bufspace -= bp->b_bufsize; 1160 bufmallocspace -= bp->b_bufsize; 1161 bp->b_data = (caddr_t) buffers_kva + (bp - buf) * MAXBSIZE; 1162 bp->b_bufsize = 0; 1163 bp->b_bcount = 0; 1164 bp->b_flags &= ~B_MALLOC; 1165 } 1166 return 1; 1167 } 1168#endif 1169 vm_hold_free_pages( 1170 bp, 1171 (vm_offset_t) bp->b_data + newbsize, 1172 (vm_offset_t) bp->b_data + bp->b_bufsize); 1173 } else if (newbsize > bp->b_bufsize) { 1174#if !defined(NO_B_MALLOC) 1175 /* 1176 * We only use malloced memory on the first allocation. 1177 * and revert to page-allocated memory when the buffer grows. 1178 */ 1179 if ( (bufmallocspace < maxbufmallocspace) && 1180 (bp->b_bufsize == 0) && 1181 (mbsize <= PAGE_SIZE/2)) { 1182 1183 bp->b_data = malloc(mbsize, M_BIOBUF, M_WAITOK); 1184 bp->b_bufsize = mbsize; 1185 bp->b_bcount = size; 1186 bp->b_flags |= B_MALLOC; 1187 bufspace += mbsize; 1188 bufmallocspace += mbsize; 1189 return 1; 1190 } 1191#endif 1192 origbuf = NULL; 1193 origbufsize = 0; 1194#if !defined(NO_B_MALLOC) 1195 /* 1196 * If the buffer is growing on it's other-than-first allocation, 1197 * then we revert to the page-allocation scheme. 1198 */ 1199 if (bp->b_flags & B_MALLOC) { 1200 origbuf = bp->b_data; 1201 origbufsize = bp->b_bufsize; 1202 bp->b_data = (caddr_t) buffers_kva + (bp - buf) * MAXBSIZE; 1203 bufspace -= bp->b_bufsize; 1204 bufmallocspace -= bp->b_bufsize; 1205 bp->b_bufsize = 0; 1206 bp->b_flags &= ~B_MALLOC; 1207 newbsize = round_page(newbsize); 1208 } 1209#endif 1210 vm_hold_load_pages( 1211 bp, 1212 (vm_offset_t) bp->b_data + bp->b_bufsize, 1213 (vm_offset_t) bp->b_data + newbsize); 1214#if !defined(NO_B_MALLOC) 1215 if (origbuf) { 1216 bcopy(origbuf, bp->b_data, origbufsize); 1217 free(origbuf, M_BIOBUF); 1218 } 1219#endif 1220 } 1221 } else { 1222 vm_page_t m; 1223 int desiredpages; 1224 1225 newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1); 1226 desiredpages = (round_page(newbsize) >> PAGE_SHIFT); 1227 1228#if !defined(NO_B_MALLOC) 1229 if (bp->b_flags & B_MALLOC) 1230 panic("allocbuf: VMIO buffer can't be malloced"); 1231#endif 1232 1233 if (newbsize < bp->b_bufsize) { 1234 if (desiredpages < bp->b_npages) { 1235 for (i = desiredpages; i < bp->b_npages; i++) { 1236 /* 1237 * the page is not freed here -- it 1238 * is the responsibility of vnode_pager_setsize 1239 */ 1240 m = bp->b_pages[i]; 1241 s = splhigh(); 1242 while ((m->flags & PG_BUSY) || (m->busy != 0)) { 1243 m->flags |= PG_WANTED; 1244 tsleep(m, PVM, "biodep", 0); 1245 } 1246 splx(s); 1247 1248 bp->b_pages[i] = NULL; 1249 vm_page_unwire(m); 1250 } 1251 pmap_qremove((vm_offset_t) trunc_page(bp->b_data) + 1252 (desiredpages << PAGE_SHIFT), (bp->b_npages - desiredpages)); 1253 bp->b_npages = desiredpages; 1254 } 1255 } else if (newbsize > bp->b_bufsize) { 1256 vm_object_t obj; 1257 vm_offset_t tinc, toff; 1258 vm_ooffset_t off; 1259 vm_pindex_t objoff; 1260 int pageindex, curbpnpages; 1261 struct vnode *vp; 1262 int bsize; 1263 1264 vp = bp->b_vp; 1265 1266 if (vp->v_type == VBLK) 1267 bsize = DEV_BSIZE; 1268 else 1269 bsize = vp->v_mount->mnt_stat.f_iosize; 1270 1271 if (bp->b_npages < desiredpages) { 1272 obj = vp->v_object; 1273 tinc = PAGE_SIZE; 1274 if (tinc > bsize) 1275 tinc = bsize; 1276 off = (vm_ooffset_t) bp->b_lblkno * bsize; 1277 doretry: 1278 curbpnpages = bp->b_npages; 1279 bp->b_flags |= B_CACHE; 1280 for (toff = 0; toff < newbsize; toff += tinc) { 1281 int bytesinpage; 1282 1283 pageindex = toff >> PAGE_SHIFT; 1284 objoff = OFF_TO_IDX(off + toff); 1285 if (pageindex < curbpnpages) { 1286 1287 m = bp->b_pages[pageindex]; 1288#ifdef VFS_BIO_DIAG 1289 if (m->pindex != objoff) 1290 panic("allocbuf: page changed offset??!!!?"); 1291#endif 1292 bytesinpage = tinc; 1293 if (tinc > (newbsize - toff)) 1294 bytesinpage = newbsize - toff; 1295 if ((bp->b_flags & B_CACHE) && 1296 !vm_page_is_valid(m, 1297 (vm_offset_t) ((toff + off) & PAGE_MASK), 1298 bytesinpage)) { 1299 bp->b_flags &= ~B_CACHE; 1300 } 1301 continue; 1302 } 1303 m = vm_page_lookup(obj, objoff); 1304 if (!m) { 1305 m = vm_page_alloc(obj, objoff, VM_ALLOC_NORMAL); 1306 if (!m) { 1307 VM_WAIT; 1308 goto doretry; 1309 } 1310 /* 1311 * Normally it is unwise to clear PG_BUSY without 1312 * PAGE_WAKEUP -- but it is okay here, as there is 1313 * no chance for blocking between here and vm_page_alloc 1314 */ 1315 m->flags &= ~PG_BUSY; 1316 vm_page_wire(m); 1317 bp->b_flags &= ~B_CACHE; 1318 } else if (m->flags & PG_BUSY) { 1319 1320 s = splhigh(); 1321 m->flags |= PG_WANTED; 1322 tsleep(m, PVM, "pgtblk", 0); 1323 splx(s); 1324 1325 goto doretry; 1326 } else { 1327 if ((curproc != pageproc) && 1328 ((m->queue - m->pc) == PQ_CACHE) && 1329 ((cnt.v_free_count + cnt.v_cache_count) < 1330 (cnt.v_free_min + cnt.v_cache_min))) { 1331 pagedaemon_wakeup(); 1332 } 1333 bytesinpage = tinc; 1334 if (tinc > (newbsize - toff)) 1335 bytesinpage = newbsize - toff; 1336 if ((bp->b_flags & B_CACHE) && 1337 !vm_page_is_valid(m, 1338 (vm_offset_t) ((toff + off) & PAGE_MASK), 1339 bytesinpage)) { 1340 bp->b_flags &= ~B_CACHE; 1341 } 1342 vm_page_wire(m); 1343 } 1344 bp->b_pages[pageindex] = m; 1345 curbpnpages = pageindex + 1; 1346 } 1347 bp->b_data = (caddr_t) trunc_page(bp->b_data); 1348 bp->b_npages = curbpnpages; 1349 pmap_qenter((vm_offset_t) bp->b_data, 1350 bp->b_pages, bp->b_npages); 1351 ((vm_offset_t) bp->b_data) |= off & PAGE_MASK; 1352 } 1353 } 1354 } 1355 if (bp->b_flags & B_VMIO) 1356 vmiospace += bp->b_bufsize; 1357 bufspace += (newbsize - bp->b_bufsize); 1358 bp->b_bufsize = newbsize; 1359 bp->b_bcount = size; 1360 return 1; 1361} 1362 1363/* 1364 * Wait for buffer I/O completion, returning error status. 1365 */ 1366int 1367biowait(register struct buf * bp) 1368{ 1369 int s; 1370 1371 s = splbio(); 1372 while ((bp->b_flags & B_DONE) == 0) 1373 tsleep(bp, PRIBIO, "biowait", 0); 1374 splx(s); 1375 if (bp->b_flags & B_EINTR) { 1376 bp->b_flags &= ~B_EINTR; 1377 return (EINTR); 1378 } 1379 if (bp->b_flags & B_ERROR) { 1380 return (bp->b_error ? bp->b_error : EIO); 1381 } else { 1382 return (0); 1383 } 1384} 1385 1386/* 1387 * Finish I/O on a buffer, calling an optional function. 1388 * This is usually called from interrupt level, so process blocking 1389 * is not *a good idea*. 1390 */ 1391void 1392biodone(register struct buf * bp) 1393{ 1394 int s; 1395 1396 s = splbio(); 1397 if (!(bp->b_flags & B_BUSY)) 1398 panic("biodone: buffer not busy"); 1399 1400 if (bp->b_flags & B_DONE) { 1401 splx(s); 1402 printf("biodone: buffer already done\n"); 1403 return; 1404 } 1405 bp->b_flags |= B_DONE; 1406 1407 if ((bp->b_flags & B_READ) == 0) { 1408 vwakeup(bp); 1409 } 1410#ifdef BOUNCE_BUFFERS 1411 if (bp->b_flags & B_BOUNCE) 1412 vm_bounce_free(bp); 1413#endif 1414 1415 /* call optional completion function if requested */ 1416 if (bp->b_flags & B_CALL) { 1417 bp->b_flags &= ~B_CALL; 1418 (*bp->b_iodone) (bp); 1419 splx(s); 1420 return; 1421 } 1422 if (bp->b_flags & B_VMIO) { 1423 int i, resid; 1424 vm_ooffset_t foff; 1425 vm_page_t m; 1426 vm_object_t obj; 1427 int iosize; 1428 struct vnode *vp = bp->b_vp; 1429 1430 if (vp->v_type == VBLK) 1431 foff = (vm_ooffset_t) DEV_BSIZE * bp->b_lblkno; 1432 else 1433 foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 1434 obj = vp->v_object; 1435 if (!obj) { 1436 panic("biodone: no object"); 1437 } 1438#if defined(VFS_BIO_DEBUG) 1439 if (obj->paging_in_progress < bp->b_npages) { 1440 printf("biodone: paging in progress(%d) < bp->b_npages(%d)\n", 1441 obj->paging_in_progress, bp->b_npages); 1442 } 1443#endif 1444 iosize = bp->b_bufsize; 1445 for (i = 0; i < bp->b_npages; i++) { 1446 int bogusflag = 0; 1447 m = bp->b_pages[i]; 1448 if (m == bogus_page) { 1449 bogusflag = 1; 1450 m = vm_page_lookup(obj, OFF_TO_IDX(foff)); 1451 if (!m) { 1452#if defined(VFS_BIO_DEBUG) 1453 printf("biodone: page disappeared\n"); 1454#endif 1455 --obj->paging_in_progress; 1456 continue; 1457 } 1458 bp->b_pages[i] = m; 1459 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 1460 } 1461#if defined(VFS_BIO_DEBUG) 1462 if (OFF_TO_IDX(foff) != m->pindex) { 1463 printf("biodone: foff(%d)/m->pindex(%d) mismatch\n", foff, m->pindex); 1464 } 1465#endif 1466 resid = IDX_TO_OFF(m->pindex + 1) - foff; 1467 if (resid > iosize) 1468 resid = iosize; 1469 /* 1470 * In the write case, the valid and clean bits are 1471 * already changed correctly, so we only need to do this 1472 * here in the read case. 1473 */ 1474 if ((bp->b_flags & B_READ) && !bogusflag && resid > 0) { 1475 vm_page_set_validclean(m, 1476 (vm_offset_t) (foff & PAGE_MASK), resid); 1477 } 1478 1479 /* 1480 * when debugging new filesystems or buffer I/O methods, this 1481 * is the most common error that pops up. if you see this, you 1482 * have not set the page busy flag correctly!!! 1483 */ 1484 if (m->busy == 0) { 1485 printf("biodone: page busy < 0, " 1486 "pindex: %d, foff: 0x(%x,%x), " 1487 "resid: %d, index: %d\n", 1488 (int) m->pindex, (int)(foff >> 32), 1489 (int) foff & 0xffffffff, resid, i); 1490 if (vp->v_type != VBLK) 1491 printf(" iosize: %ld, lblkno: %d, flags: 0x%lx, npages: %d\n", 1492 bp->b_vp->v_mount->mnt_stat.f_iosize, 1493 (int) bp->b_lblkno, 1494 bp->b_flags, bp->b_npages); 1495 else 1496 printf(" VDEV, lblkno: %d, flags: 0x%lx, npages: %d\n", 1497 (int) bp->b_lblkno, 1498 bp->b_flags, bp->b_npages); 1499 printf(" valid: 0x%x, dirty: 0x%x, wired: %d\n", 1500 m->valid, m->dirty, m->wire_count); 1501 panic("biodone: page busy < 0\n"); 1502 } 1503 --m->busy; 1504 if ((m->busy == 0) && (m->flags & PG_WANTED)) { 1505 m->flags &= ~PG_WANTED; 1506 wakeup(m); 1507 } 1508 --obj->paging_in_progress; 1509 foff += resid; 1510 iosize -= resid; 1511 } 1512 if (obj && obj->paging_in_progress == 0 && 1513 (obj->flags & OBJ_PIPWNT)) { 1514 obj->flags &= ~OBJ_PIPWNT; 1515 wakeup(obj); 1516 } 1517 } 1518 /* 1519 * For asynchronous completions, release the buffer now. The brelse 1520 * checks for B_WANTED and will do the wakeup there if necessary - so 1521 * no need to do a wakeup here in the async case. 1522 */ 1523 1524 if (bp->b_flags & B_ASYNC) { 1525 if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR | B_RELBUF)) != 0) 1526 brelse(bp); 1527 else 1528 bqrelse(bp); 1529 } else { 1530 wakeup(bp); 1531 } 1532 splx(s); 1533} 1534 1535int 1536count_lock_queue() 1537{ 1538 int count; 1539 struct buf *bp; 1540 1541 count = 0; 1542 for (bp = TAILQ_FIRST(&bufqueues[QUEUE_LOCKED]); 1543 bp != NULL; 1544 bp = TAILQ_NEXT(bp, b_freelist)) 1545 count++; 1546 return (count); 1547} 1548 1549int vfs_update_interval = 30; 1550 1551static void 1552vfs_update() 1553{ 1554 (void) spl0(); /* XXX redundant? wrong place? */ 1555 while (1) { 1556 tsleep(&vfs_update_wakeup, PUSER, "update", 1557 hz * vfs_update_interval); 1558 vfs_update_wakeup = 0; 1559 sync(curproc, NULL, NULL); 1560 } 1561} 1562 1563static int 1564sysctl_kern_updateinterval SYSCTL_HANDLER_ARGS 1565{ 1566 int error = sysctl_handle_int(oidp, 1567 oidp->oid_arg1, oidp->oid_arg2, req); 1568 if (!error) 1569 wakeup(&vfs_update_wakeup); 1570 return error; 1571} 1572 1573SYSCTL_PROC(_kern, KERN_UPDATEINTERVAL, update, CTLTYPE_INT|CTLFLAG_RW, 1574 &vfs_update_interval, 0, sysctl_kern_updateinterval, "I", ""); 1575 1576 1577/* 1578 * This routine is called in lieu of iodone in the case of 1579 * incomplete I/O. This keeps the busy status for pages 1580 * consistant. 1581 */ 1582void 1583vfs_unbusy_pages(struct buf * bp) 1584{ 1585 int i; 1586 1587 if (bp->b_flags & B_VMIO) { 1588 struct vnode *vp = bp->b_vp; 1589 vm_object_t obj = vp->v_object; 1590 vm_ooffset_t foff; 1591 1592 foff = (vm_ooffset_t) vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 1593 1594 for (i = 0; i < bp->b_npages; i++) { 1595 vm_page_t m = bp->b_pages[i]; 1596 1597 if (m == bogus_page) { 1598 m = vm_page_lookup(obj, OFF_TO_IDX(foff) + i); 1599 if (!m) { 1600 panic("vfs_unbusy_pages: page missing\n"); 1601 } 1602 bp->b_pages[i] = m; 1603 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 1604 } 1605 --obj->paging_in_progress; 1606 --m->busy; 1607 if ((m->busy == 0) && (m->flags & PG_WANTED)) { 1608 m->flags &= ~PG_WANTED; 1609 wakeup(m); 1610 } 1611 } 1612 if (obj->paging_in_progress == 0 && 1613 (obj->flags & OBJ_PIPWNT)) { 1614 obj->flags &= ~OBJ_PIPWNT; 1615 wakeup(obj); 1616 } 1617 } 1618} 1619 1620/* 1621 * This routine is called before a device strategy routine. 1622 * It is used to tell the VM system that paging I/O is in 1623 * progress, and treat the pages associated with the buffer 1624 * almost as being PG_BUSY. Also the object paging_in_progress 1625 * flag is handled to make sure that the object doesn't become 1626 * inconsistant. 1627 */ 1628void 1629vfs_busy_pages(struct buf * bp, int clear_modify) 1630{ 1631 int i; 1632 1633 if (bp->b_flags & B_VMIO) { 1634 vm_object_t obj = bp->b_vp->v_object; 1635 vm_ooffset_t foff; 1636 int iocount = bp->b_bufsize; 1637 1638 if (bp->b_vp->v_type == VBLK) 1639 foff = (vm_ooffset_t) DEV_BSIZE * bp->b_lblkno; 1640 else 1641 foff = (vm_ooffset_t) bp->b_vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 1642 vfs_setdirty(bp); 1643 for (i = 0; i < bp->b_npages; i++) { 1644 vm_page_t m = bp->b_pages[i]; 1645 int resid = IDX_TO_OFF(m->pindex + 1) - foff; 1646 1647 if (resid > iocount) 1648 resid = iocount; 1649 if ((bp->b_flags & B_CLUSTER) == 0) { 1650 obj->paging_in_progress++; 1651 m->busy++; 1652 } 1653 vm_page_protect(m, VM_PROT_NONE); 1654 if (clear_modify) { 1655 vm_page_set_validclean(m, 1656 (vm_offset_t) (foff & PAGE_MASK), resid); 1657 } else if (bp->b_bcount >= PAGE_SIZE) { 1658 if (m->valid && (bp->b_flags & B_CACHE) == 0) { 1659 bp->b_pages[i] = bogus_page; 1660 pmap_qenter(trunc_page(bp->b_data), bp->b_pages, bp->b_npages); 1661 } 1662 } 1663 foff += resid; 1664 iocount -= resid; 1665 } 1666 } 1667} 1668 1669/* 1670 * Tell the VM system that the pages associated with this buffer 1671 * are clean. This is used for delayed writes where the data is 1672 * going to go to disk eventually without additional VM intevention. 1673 */ 1674void 1675vfs_clean_pages(struct buf * bp) 1676{ 1677 int i; 1678 1679 if (bp->b_flags & B_VMIO) { 1680 vm_ooffset_t foff; 1681 int iocount = bp->b_bufsize; 1682 1683 if (bp->b_vp->v_type == VBLK) 1684 foff = (vm_ooffset_t) DEV_BSIZE * bp->b_lblkno; 1685 else 1686 foff = (vm_ooffset_t) bp->b_vp->v_mount->mnt_stat.f_iosize * bp->b_lblkno; 1687 1688 for (i = 0; i < bp->b_npages; i++) { 1689 vm_page_t m = bp->b_pages[i]; 1690 int resid = IDX_TO_OFF(m->pindex + 1) - foff; 1691 1692 if (resid > iocount) 1693 resid = iocount; 1694 if (resid > 0) { 1695 vm_page_set_validclean(m, 1696 ((vm_offset_t) foff & PAGE_MASK), resid); 1697 } 1698 foff += resid; 1699 iocount -= resid; 1700 } 1701 } 1702} 1703 1704void 1705vfs_bio_clrbuf(struct buf *bp) { 1706 int i; 1707 if( bp->b_flags & B_VMIO) { 1708 if( (bp->b_npages == 1) && (bp->b_bufsize < PAGE_SIZE)) { 1709 int mask; 1710 mask = 0; 1711 for(i=0;i<bp->b_bufsize;i+=DEV_BSIZE) 1712 mask |= (1 << (i/DEV_BSIZE)); 1713 if( bp->b_pages[0]->valid != mask) { 1714 bzero(bp->b_data, bp->b_bufsize); 1715 } 1716 bp->b_pages[0]->valid = mask; 1717 bp->b_resid = 0; 1718 return; 1719 } 1720 for(i=0;i<bp->b_npages;i++) { 1721 if( bp->b_pages[i]->valid == VM_PAGE_BITS_ALL) 1722 continue; 1723 if( bp->b_pages[i]->valid == 0) { 1724 if ((bp->b_pages[i]->flags & PG_ZERO) == 0) { 1725 bzero(bp->b_data + (i << PAGE_SHIFT), PAGE_SIZE); 1726 } 1727 } else { 1728 int j; 1729 for(j=0;j<PAGE_SIZE/DEV_BSIZE;j++) { 1730 if( (bp->b_pages[i]->valid & (1<<j)) == 0) 1731 bzero(bp->b_data + (i << PAGE_SHIFT) + j * DEV_BSIZE, DEV_BSIZE); 1732 } 1733 } 1734 /* bp->b_pages[i]->valid = VM_PAGE_BITS_ALL; */ 1735 } 1736 bp->b_resid = 0; 1737 } else { 1738 clrbuf(bp); 1739 } 1740} 1741 1742/* 1743 * vm_hold_load_pages and vm_hold_unload pages get pages into 1744 * a buffers address space. The pages are anonymous and are 1745 * not associated with a file object. 1746 */ 1747void 1748vm_hold_load_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 1749{ 1750 vm_offset_t pg; 1751 vm_page_t p; 1752 int index; 1753 1754 to = round_page(to); 1755 from = round_page(from); 1756 index = (from - trunc_page(bp->b_data)) >> PAGE_SHIFT; 1757 1758 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 1759 1760tryagain: 1761 1762 p = vm_page_alloc(kernel_object, ((pg - VM_MIN_KERNEL_ADDRESS) >> PAGE_SHIFT), 1763 VM_ALLOC_NORMAL); 1764 if (!p) { 1765 VM_WAIT; 1766 goto tryagain; 1767 } 1768 vm_page_wire(p); 1769 pmap_kenter(pg, VM_PAGE_TO_PHYS(p)); 1770 bp->b_pages[index] = p; 1771 PAGE_WAKEUP(p); 1772 } 1773 bp->b_npages = to >> PAGE_SHIFT; 1774} 1775 1776void 1777vm_hold_free_pages(struct buf * bp, vm_offset_t from, vm_offset_t to) 1778{ 1779 vm_offset_t pg; 1780 vm_page_t p; 1781 int index; 1782 1783 from = round_page(from); 1784 to = round_page(to); 1785 index = (from - trunc_page(bp->b_data)) >> PAGE_SHIFT; 1786 1787 for (pg = from; pg < to; pg += PAGE_SIZE, index++) { 1788 p = bp->b_pages[index]; 1789 if (p && (index < bp->b_npages)) { 1790 if (p->busy) { 1791 printf("vm_hold_free_pages: blkno: %d, lblkno: %d\n", 1792 bp->b_blkno, bp->b_lblkno); 1793 } 1794 bp->b_pages[index] = NULL; 1795 pmap_kremove(pg); 1796 vm_page_unwire(p); 1797 vm_page_free(p); 1798 } 1799 } 1800 bp->b_npages = from >> PAGE_SHIFT; 1801} 1802